Some checks failed
CI / lint (push) Successful in 1m26s
CI / notebooks-smoke (push) Successful in 3m33s
Deploy / changes (push) Successful in 50s
CI / test (push) Successful in 5m46s
Infra CI / notebooks (push) Successful in 1m13s
Infra CI / zotero (push) Successful in 23s
Infra CI / docs (push) Successful in 39s
Infra CI / api (push) Successful in 1m47s
Infra CI / bcda (push) Successful in 2m27s
Infra CI / llm (push) Successful in 1m28s
Infra CI / mc (push) Successful in 27s
Deploy / zotero (push) Has been skipped
Deploy / docs (push) Successful in 1m8s
Deploy / report (push) Has been cancelled
Deploy / llm (push) Has been cancelled
Deploy / mc (push) Has been cancelled
Deploy / bcda (push) Has been cancelled
Deploy / notebooks (push) Has been cancelled
Deploy / api (push) Has been cancelled
The prompt was built from each source's 500-char UI snippet, so the model never saw anything past it — the CY2027 final CF sentence starts at char 518 of its paragraph. Sources now carry the chunk text (capped at 1600 chars) for the prompt; the snippet stays for the UI. And a rule paragraph hit carries the paragraph after it: FR lead-ins put the figures in the next paragraph, which retrieval often misses. Asked for the CY2027 final QP/non-QP CFs, chat answered the NPRM's $33.17/$32.84 from a public comment; it now answers 32.9548/32.6286 from 91 FR 71653.